{"id":"b02de435-bc66-4b54-b356-187812aa8eae","arxiv_id":"2506.19890","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A causal-reinforcement-learning resource allocator for VR interactions, PS-CDDPG, jointly selects keyframe ratios, bandwidth, and CPU frequency and is reported to cut training iterations by over 30 percent while raising a Weber-Fechner-law-based QoE metric.","lead":"This paper builds an AI controller that chooses how many animation keyframes, how much bandwidth, and how much computing power each user gets in a multi-user virtual reality session. It claims the controller, which uses causal reasoning about which actions actually affect user experience, trains faster and improves a designed quality-of-experience score compared to plain deep reinforcement learning.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Causal component is never isolated: the 30% training gain may come from sampling 64 candidate actions, not from the CAI ranking.","rationale":"The reader's weakest_assumption concerns the accuracy of the Gaussian transition model. My concern is one step upstream: even if the Gaussian model were accurate, the experiments never prove that the CAI ranking, rather than the enlarged candidate set, causes the faster convergence. This is the load-bearing link between the algorithm mechanism and the headline claim. The proposed test is a minimal ablation: it changes only the selection rule, holding the candidate-generation distribution and all training hyperparameters fixed. It also addresses the single-seed issue by requiring at least 10 seeds and confidence intervals. If the ablation shows no difference, the causal-aware contribution is not established and the paper should be revised accordingly; if it shows a clear difference, the central claim gains support, though the Gaussian approximation concern for real deployment remains. I therefore recommend keeping the reader's CONDITIONAL verdict rather than moving to reject or accept.","tokens_in":19490,"tokens_out":10666,"duration_ms":118417,"concrete_test":"Run PS-CDDPG with the CAI-based selection in Eqs. (21)-(22) replaced by uniform random selection among the same N=64 noise candidates, keeping epsilon=0.4, sigma_eta^2=0.01, and all other hyperparameters identical. Use at least 10 seeds and report mean and confidence intervals for iterations to reach average reward 1.45. If the random-selection variant matches PS-CDDPG, the causal scores are not load-bearing; if it degrades to DDPG-level, the causal ranking is doing the work.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim (Section VI) attributes a >30% reduction in training iterations to causal influence detection guiding exploration. The comparison in Section V-B does not isolate the causal information: PS-CDDPG (Algorithm 2) generates N=64 noisy candidate actions and selects among them using CAI scores, while the DDPG baseline generates one noisy action. Thus the observed gain conflates two changes: using a larger candidate set before acting, and ranking that set by CAI. The CAI scores themselves are produced by a Gaussian transition model (Eqs. 15-18) that the authors concede is a limited approximation in complex environments (Section VI). If CAI ranking were replaced by random selection from the same candidate set, exploration might still improve and the stated causal mechanism would not be the source of the gain. No ablation with randomized CAI scores is reported; without it, the causal explanation of the 30% improvement is unverified.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a framework for QoE optimization in multi-user VR interaction with adaptive keyframe extraction. It defines a QoE metric based on the Weber-Fechner law combined with attention levels and keyframe ratios, then formulates the joint allocation of keyframe ratio, bandwidth, and CPU frequency as a mixed-integer optimization problem under fairness constraints. The proposed algorithm, PS-CDDPG, augments DDPG with causal influence detection (CAI), partial state division, and noise-based active exploration. Experiments using the CMU Motion Capture Database simulate a sub-6 GHz multi-user environment and report faster convergence, reduced latency, improved QoE, and fairness relative to several baselines.","tokens_in":19677,"tokens_out":5915,"duration_ms":63713,"significance":"If the central claims hold, the paper provides a useful application of causal influence detection to wireless VR resource allocation and offers a complete algorithm description with reproducible public-data simulation. Strengths include the explicit MIP formulation, the detailed PS-CDDPG pseudocode, and the sensitivity analysis of exploration hyperparameters. However, the causal mechanism is not isolated from the enlarged candidate action set, the Gaussian transition assumption for CAI is not validated, and the reported benefits lack statistical uncertainty. These gaps currently prevent the claimed causal contribution from being accepted as stated.","major_comments":[{"comment":"The central claim that causal influence detection yields a >30% reduction in training iterations is not isolated from the enlarged candidate set. PS-CDDPG (Algorithm 2, lines 10-18) generates N=64 noisy candidate actions and selects among them using CAI weights, whereas the DDPG baseline perturbs the actor action with a single noise vector. The convergence comparisons in Figs. 3-4 therefore vary two factors at once: candidate-set size and the selection criterion. Without an ablation that keeps the 64-candidate set but replaces CAI ranking with random, reverse, or permuted CAI scores, the reported training gain cannot be attributed to causal influence detection rather than to the benefit of evaluating multiple candidate actions. This ablation is needed to support the paper's primary contribution.","section":"Section V-B, Fig. 3, Algorithm 2"},{"comment":"The CAI score assumes that each next-state variable is Gaussian given state and action, with mean and variance produced by a DNN. The authors concede in Section VI that the DNN inference model is inaccurate in complex environments, but no diagnostic is provided for the quality of the fitted transition distribution, such as calibration curves, comparison with non-parametric estimates, or checks on the KL divergence used in Eq. (17). If the Gaussian approximation is poor, CAI scores may rank actions according to approximation error rather than true causal influence, and the exploration guidance would be noise-driven rather than causal. The causal explanation of the training gain is therefore unsupported without an inference-model validation or an ablation using an alternative transition model.","section":"Section IV-B, Eqs. (15)-(18); Section VI"},{"comment":"The evaluation reports single training curves and test averages without confidence intervals, multiple random seeds, or significance tests. The claims that the method 'significantly reduces latency', 'enhances QoE', and 'outperforms baseline approaches' require statistical support; please report means and variances over independent seeds or dataset subsamples, and where appropriate significance tests, for the main comparisons in Figs. 3, 5, 6, and 12.","section":"Section V, Figs. 3-12"},{"comment":"The QoE metric in Eq. (8) is an analytical construction combining attention weights, latency, and a logarithmic keyframe term, and this same expression is used as the reward in Eq. (14). No external validation against subjective QoE scores or established QoE instruments is provided. Because the metric itself encodes a preference for attention-weighted keyframe allocation, methods that follow this preference will mechanically score higher. The paper should either validate the QoE proxy or explicitly frame the contribution as optimizing a proposed analytical metric rather than a measured user experience.","section":"Section III-D, Eq. (8); Section IV-A, Eq. (14)"}],"minor_comments":[{"comment":"The baseline list includes an 'Adaptive Keyframe' method, but the text and figures treat 'Adaptive Keyframe' as the proposed PS-CDDPG method. Please rename the proposed method (e.g., 'PS-CDDPG') and clarify whether the bullet-list 'Adaptive Keyframe' is a baseline or the proposed framework.","section":"Section V-D, Figs. 6-12"},{"comment":"The exploration rule is ambiguous: Eq. (21) selects the candidate with the largest CAI score, whereas Eq. (22) defines selection probabilities by rank, and Algorithm 2 only refers to 'calculate priority weights' and 'select action'. Please state explicitly whether the exploration choice is deterministic argmax or stochastic over the weights, and where Eq. (22) is applied.","section":"Section IV-C, Eqs. (21)-(22), Algorithm 2"},{"comment":"The rendering CPU frequency f_r_k(t) is given as [1.5-2.5] GHz in Table II, but the optimization variables and Algorithm 2 only allocate f_e_k(t). Please clarify how f_r_k(t) is set or sampled in the simulation.","section":"Section V-A, Table II"},{"comment":"The partition of the state into S^(1)(t) and S^(2)(t) is described informally. Please provide a formal definition of these two sets and explain how S^(2)(t+1) is computed or observed in the environment step.","section":"Section IV-C, Eq. (12), Algorithm 2"},{"comment":"Panels (c)-(f) plot multiple curves for different epsilon and sigma values without distinct markers and with unclear legend placement, making the comparisons difficult to read. Please improve the figure with explicit legends or separate subplots.","section":"Fig. 5"}],"recommendation":"major_revision","confidential_remarks":"The paper is within scope for a communications/multimedia journal, and the simulation is carefully parameterized. The main risk is overclaiming causality: the requested ablation and seed-level statistics are essential before the central contribution can be credited."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First thing to know: this is a competent, clearly written engineering paper, but its central claim—that causal influence detection cuts training iterations by 30%—is not actually isolated. PS-CDDPG samples 64 noise-perturbed candidate actions and picks one via CAI scores; the DDPG baseline samples one noisy action. So the comparison conflates the larger candidate set with causal ranking. There is no ablation with random selection from the same 64 candidates, which is exactly the control needed. The stress-test note lands.\n\nWhat is new: the combination of partial-state causal influence detection (from Seitzer et al. and Wang et al.) with a continuous-action noise-based active exploration, applied to a mixed-integer VR resource allocation problem. The system model is detailed, the algorithm pseudocode is complete, and the evaluation uses CMU motion capture data. The paper is honest about the Gaussian transition assumption being a limitation in complex environments.\n\nThe state division itself does get some support: PS-CDDPG is compared against CAI+DDPG (same candidate set and CAI ranking, no state division), and PS-CDDPG converges faster. So the improvement over CAI+DDPG is attributable to partial state abstraction. That is a real, if modest, isolated result.\n\nWhere it is soft: (1) No error bars or multiple seeds; all curves appear to come from a single run. (2) The 'Adaptive Keyframe' baseline in Section V-D is ambiguous—it could be a rule-based heuristic or the proposed method itself, and the reader should not have to guess. (3) The QoE metric is a customized formula, never validated against user studies or an accepted QoE model; the reward is built from it, so optimizing reward means optimizing the authors' metric. This is a design choice, not circularity, but it limits external inference. (4) The 30% claim is a bit underspecified: what is the 'desired average reward'?\n\nProportionately, these are fixable issues. The core framework is plausible and the paper is reproducible enough to be worth referee time. I recommend a serious peer review with a request for the missing ablation, multiple seeds, and baseline clarification. If random ranking from the same 64 candidates does not recreate the gain, the causal story will be much stronger.","headline":"Solid engineering paper whose headline causal speedup claim lacks a control for the larger action candidate set; worth reviewing but needs an ablation.","tokens_in":20184,"tokens_out":4518,"would_cite":false,"duration_ms":41508,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Causal-aware exploration cuts VR resource-allocation training by 30 percent.","keywords":["quality of experience","virtual reality","keyframe extraction","causal reinforcement learning","DDPG","resource allocation","Weber-Fechner Law","mixed integer programming"],"falsifier":"Reproduce the training procedure on an environment with the same state-action structure but highly non-Gaussian next-state distributions (for example, multimodal or heavy-tailed latency), and check whether PS-CDDPG still reaches the target average reward in 30 percent fewer iterations than DDPG; if the gain vanishes or becomes negative, the Gaussian inference model is the load-bearing assumption.","tokens_in":19279,"feed_emoji":"🎮","tokens_out":3521,"duration_ms":29740,"temperature":0.7,"pith_summary":"This paper argues that a multi-user VR system can serve more users with better perceived quality if the resource allocator uses causal information, not just trial-and-error, to guide its learning. It builds a QoE metric from the Weber-Fechner Law, ties it to attention levels and keyframe-based motion transmission, and then introduces PS-CDDPG, a DDPG variant whose exploration is steered by inferred causal influence scores. The authors claim the framework reaches the target average reward with over 30 percent fewer training iterations, and that in deployment it lowers latency, raises QoE, and preserves fairness relative to fixed-ratio and attention-only baselines. If right, the main payoff is practical: causal structure extracted during training can make RL-based network resource allocation faster to deploy and more stable in dynamic settings.","feed_headline":"Causal-aware exploration cuts VR resource-allocation training by 30 percent","feed_subtitle":"A DDPG variant that ranks exploratory actions by inferred causal influence trains faster and serves multi-user VR more fairly.","key_machinery":"The causal action influence (CAI) score is the central object: a per-next-state-variable conditional mutual information $I(S'_j;A|S=s,A=a)$, computed as the KL divergence between the action-conditioned transition distribution $p(s'_j|s,a)$ and the action-marginal $p(s'_j|s)$, averaged over $|S'|$ next-state variables and approximated by a Monte-Carlo average over sampled actions. A DNN inference model, trained by negative log-likelihood under a Gaussian assumption $S'_j \\sim \\mathcal{N}(\\mu_\\theta(s,a), \\sigma_\\theta^2(s,a))$, supplies the distributions; the score's job is to rank candidate exploratory actions so that the agent preferentially tries actions that strongly control the next state. The second machinery piece is the partial-state division, which restricts CAI computation to genuinely action-relevant variables so the learned inference model does not waste capacity on attention data that the actions cannot influence.","core_discovery":"The central claim is that causal influence detection, applied only to action-relevant state variables and combined with noise-based active exploration, makes Deep Deterministic Policy Gradient training substantially more efficient for a mixed-integer QoE optimization problem in multi-user VR interaction. The paper defines a QoE function $QoE_k(t)$ based on the Weber-Fechner Law that combines latency ratio, attention-weighted character counts, and a logarithmic keyframe term, and models the joint optimization of keyframe ratio, bandwidth, and CPU frequency as a constrained MIP. To solve it, the state is split into action-relevant variables $S^{(2)}(t)$ and action-irrelevant variables $S^{(1)}(t)$ (user attention data), and a neural inference model approximates the transition distributions $p(s'_j|s,a)$ as Gaussians whose KL divergence yields the causal action influence (CAI) score. Exploration then samples a batch of noise-perturbed actions and selects among them using weight-based CAI ranking. The paper reports that the full PS-CDDPG framework reaches the desired average reward in over 30 percent fewer iterations than plain DDPG or CAI+DDPG without the state split, and that in the test environment it reduces latency, improves QoE, and maintains horizon fairness against several baselines.","pith_inferences":["The paper's own sensitivity analysis suggests an implicit test: the 30 percent iteration reduction should be largest when the Gaussian inference model is accurate; environments with strongly non-Gaussian transition dynamics should shrink or erase the gain, since the CAI ranking then approximates random noise.","A neighboring application of the same machinery is multi-user cloud gaming or AR, where bandwidth, edge CPU, and frame-selection decisions share the same mixed-integer structure; the partial-state CAI scheme could be ported with the attention map replaced by gaze or saliency data.","The weight-based selection in (22) could be made adaptive: annealing exploration noise variance $\\sigma_\\eta^2$ or the exploration rate $\\epsilon$ according to the agreement between CAI-ranked actions and observed reward would test whether the causal signal remains informative after convergence.","A stricter causal standard would replace the Gaussian KL proxy with a nonparametric conditional-dependence test on the same replay data; if the PS-CDDPG gain persists under that replacement, the mechanism is causal structure rather than distributional filtering."],"forward_implications":["If the reported training gain holds across environments, causal-influence-guided exploration becomes a drop-in accelerant for DDPG-style allocation policies in communication networks, since it modifies only the exploration phase.","The attention-based keyframe model implies that transmitting fewer frames for low-attention characters and more for central vision can keep QoE high while cutting bandwidth demand; the paper shows the adaptive scheme outperforms fixed keyframe ratios, especially at limited bandwidth.","Splitting state variables into action-relevant and action-irrelevant sets should generalize to other RL resource allocation tasks where some inputs, like user position or channel randomness, are outside the agent's control.","The horizon-fairness constraint embedded in the reward suggests the causal-exploration approach can trade off average QoE and per-user variance without a separate fairness-training stage."],"supporting_citations":[{"why":"Supplies the original causal influence detection method (CMI-based CAI scores, active exploration) that PS-CDDPG extends with partial state division and noise-based exploration.","marker":"[4]"},{"why":"Provides the Weber-Fechner Law basis used to construct the logarithmic QoE metric.","marker":"[21]"},{"why":"Supplies the attention-aware, sub-6 GHz edge-computing VR model and attention-based QoE formulation that this paper adapts to multi-user motion interaction.","marker":"[3]"},{"why":"Establishes that reconstruction accuracy depends primarily on keyframe ratio, justifying the keyframe-ratio decision variable and the 10 percent accuracy claim.","marker":"[2]"},{"why":"Demonstrates RL-based keyframe extraction for motion capture data, used as evidence that learned keyframe selection is feasible and that reconstruction quality rises with keyframe count.","marker":"[12]"},{"why":"Motivates the partial-state abstraction idea (task-independent state abstraction) that the paper reinterprets as splitting action-relevant from action-irrelevant state variables.","marker":"[37]"},{"why":"Provides the sub-6 GHz indoor channel and path-loss model used in the communication latency calculation.","marker":"[31]"},{"why":"Supplies the horizon-fairness QoE formulation used in the fairness constraint and reward.","marker":"[32]"}],"fun_headline_variants":["Causal-aware exploration speeds VR QoE training by 30%","Causal influence detection accelerates VR QoE RL by 30%","PS-CDDPG: causal-guided exploration cuts VR training steps 30%","Ranking actions by cause slashes VR QoE training 30%","VR QoE: causal-aware RL reaches target in 30% fewer iterations"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole training-efficiency gain rests on the assumption that the neural network's Gaussian approximation of next-state probabilities is accurate enough that the KL-divergence ranking of candidate actions reflects true causal influence; if that approximation is off, exploration is guided by noise and the reported 30 percent iteration reduction would not be explained.","fun_headline_variants_meta":{"raw":{"variants":["Causal-aware exploration speeds VR QoE training by 30%","Causal influence detection accelerates VR QoE RL by 30%","PS-CDDPG: causal-guided exploration cuts VR training steps 30%","Ranking actions by cause slashes VR QoE training 30%","VR QoE: causal-aware RL reaches target in 30% fewer iterations"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000765,"raw_usage":{"total_tokens":3452,"prompt_tokens":1064,"completion_tokens":2388,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":680,"completion_tokens_details":{"reasoning_tokens":2288}},"tokens_in":680,"tokens_out":2388,"duration_ms":16017,"temperature":1.0,"reasoning_tokens":2288,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T23:07:02.914510+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Reproduce the training procedure on an environment with the same state-action structure but highly non-Gaussian next-state distributions (for example, multimodal or heavy-tailed latency), and check whether PS-CDDPG still reaches the target average reward in 30 percent fewer iterations than DDPG; if the gain vanishes or becomes negative, the Gaussian inference model is the load-bearing assumption.","supporting_citations":[{"cited_title":"Causal influence detection for improving efficiency in reinforcement learning,","cited_arxiv_id":null,"evidence_quote":"Supplies the original causal influence detection method (CMI-based CAI scores, active exploration) that PS-CDDPG extends with partial state division and noise-based exploration."},{"cited_title":"The logarithmic nature of QoE and the role of the Weber-Fechner law in QoE assessment,","cited_arxiv_id":null,"evidence_quote":"Provides the Weber-Fechner Law basis used to construct the logarithmic QoE metric."},{"cited_title":"Attention-based QoE-aware digital twin empowered edge computing for immersive virtual reality,","cited_arxiv_id":null,"evidence_quote":"Supplies the attention-aware, sub-6 GHz edge-computing VR model and attention-based QoE formulation that this paper adapts to multi-user motion interaction."},{"cited_title":"Keyframe extraction for human motion capture data based on joint kernel sparse representation,","cited_arxiv_id":null,"evidence_quote":"Establishes that reconstruction accuracy depends primarily on keyframe ratio, justifying the keyframe-ratio decision variable and the 10 percent accuracy claim."},{"cited_title":"Keyframe extraction from motion capture sequences with graph based deep reinforcement learning,","cited_arxiv_id":null,"evidence_quote":"Demonstrates RL-based keyframe extraction for motion capture data, used as evidence that learned keyframe selection is feasible and that reconstruction quality rises with keyframe count."},{"cited_title":"Causal dynamics learning for task-independent state abstraction,","cited_arxiv_id":null,"evidence_quote":"Motivates the partial-state abstraction idea (task-independent state abstraction) that the paper reinterprets as splitting action-relevant from action-irrelevant state variables."},{"cited_title":"Sub-6 GHz channel modeling and evaluation in indoor industrial environments,","cited_arxiv_id":null,"evidence_quote":"Provides the sub-6 GHz indoor channel and path-loss model used in the communication latency calculation."},{"cited_title":"Enabling long-term fairness in dynamic resource allocation,","cited_arxiv_id":null,"evidence_quote":"Supplies the horizon-fairness QoE formulation used in the fairness constraint and reward."}],"review_version":1}